Rust: Implement type inference for closures and calls to closures #20130
Conversation
There was a problem hiding this comment.
Pull Request Overview
This PR implements type inference for closure expressions and calls to closures in Rust, including proper handling of FnOnce trait bounds. The implementation reuses tuple types for function parameters and represents closures as dyn FnOnce types.
- Adds type inference support for closure expressions, their parameters, and return types
- Implements type propagation for calls to closures and higher-order functions
- Introduces handling of the
FnOncetrait with syntactic sugar for function types
Reviewed Changes
Copilot reviewed 6 out of 6 changed files in this pull request and generated 2 comments.
Show a summary per file
| File | Description |
|---|---|
| closure.rs | New test file with comprehensive closure type inference test cases |
| main.rs | Removes old closure test module and adds reference to new test file |
| TypeMention.qll | Adds support for parenthesized argument lists and FnOnce trait handling |
| TypeInference.qll | Implements core closure type inference logic and call expression handling |
| Stdlib.qll | Defines the FnOnce trait class for type system integration |
| type-inference.expected | Expected test output showing inferred types for closure expressions |
| mod simple_closures { | ||
| pub fn test() { | ||
| // A simple closure without type annotations or invocations. | ||
| let my_closure = |a, b| a && b; |
There was a problem hiding this comment.
Does this test case show us anything?
There was a problem hiding this comment.
It shows (in the expected file) what types will be inferred for a simple closure that isn't called.
There was a problem hiding this comment.
Ah, shame it doesn't show up in an annotation, but 👍.
| @@ -2459,46 +2459,7 @@ pub mod pattern_matching_experimental { | |||
| } | |||
| } | |||
|
|
|||
| mod closures { | |||
There was a problem hiding this comment.
Why have these test cases been deleted?
There was a problem hiding this comment.
I think the stuff being tested by these examples are covered by the new tests for closures. There was quite a lot of "extra" setup in these tests, for what was actually being tested, so I think the new tests are preferable.
There was a problem hiding this comment.
Yes they were somewhat close to some real world examples I'd encountered. I guess I'd like reassuring that they would indeed work right now, whether or not they're part of our test suite.
There was a problem hiding this comment.
Sorry, I should have been more clear. They unfortunately still don't work, due to the "Another reason" that I mention in #20130 (comment).
The case is covered by the
let f = |x| x + 1; // $ MISSING: type=x:i64 target=add
let _r2 = apply_two(f); // $ target=apply_two type=_r2:i64test, which is essentially a "minimal" variant of the existing test. It test passing an unannotated closure to a function with a known type.
| ) | ||
| or | ||
| // _If_ the invoked expression has the type of a closure, then we propagate | ||
| // the surrounding types into the closure. |
There was a problem hiding this comment.
I'd like to understand how these differs from the cases above a bit better.
There was a problem hiding this comment.
The key difference is the direction.
- The cases above propagate type information out from the invoked expression (which implements
FnOnce) to the arguments and the entire call. - The cases below propagate type information in to the invoked expression. This however we only do if the invoked expression has a closure type (isn't some other type that implements
FnOnce).
Propagating type information into the invoked expression for arbitrary FnOnce implementations would make sense, but we don't have that machinery as of right now.
There was a problem hiding this comment.
Great, thanks for the explanation.
Co-authored-by: Geoffrey White <40627776+geoffw0@users.noreply.github.com>
My guess is that a big reason is that we only support Another reason could be that we can't infer the type of a closure from it being passed to an annotated function. This is shown on line 62 in the test and is an instance of a more general problem: When an argument is passed to a function where the function expects the argument to implement some trait, we don't propagate information from the trait bound back to the argument. This affects closures a lot since they are always given as trait bounds on parameter types. |
This PR implements type inference for closure expressions, calls to closures, and correctly handles
FnOncetrait bounds.A few notes on the implementation and the PR:
FnOncetrait has a type parameterArgswhich is a tuple of all the arguments. That is reflected in our implementation as well, i.e., we re-use our tuple types for parameters.FnOnce,FnMut, andFntraits. They form a trait hierarchy withFnOnceat the top.FnOnceuses an associated type for the return type and since we don't support associated types inherited in subtraits (tests for this was added in Rust: Fix type inference for trait objects for traits with associated types #20122) we can right now only really understand theFnOncetrait and not the others.dyn FnOncetype. That is, instead of creating some new type for closures we just reusedyn. The Rust language itself says that closures have some internal type that can't be written by users, and all that known about it is that it implements some of theFnOnce/Fn/FnMuttraits.dyngives us this effect. rust-analyzer shows closures as having inimpl FnOncetype. I think that makes a bit more sense thandyn, but usingdynwas simpler.Fntrait for closures, as it is a subtrait of the others, but, again, we can't understandFnyet.DCA seems fine, though we only get a 0.04% point increase in resolved calls, which is less than I had hoped.